This will greatly simplify using the common losses under sklearn._loss in (old) gradient boosting.

Meaning, this improvement will have to wait until 1.3 (PR will be large enough without keeping backwards compat loss_)😬

What do we use instead of loss_?

For most of them, your can use the equivalent function under sklearn.metrics. It also depends on your use case. What do you want to achieve?

For most of them, your can use the equivalent function under sklearn.metrics.

1. It would be nice if it were clearly documented - e.g., figuring out the need for sigmoid is not obvious.
2. Which metrics correspond to deviance and exponential?

It also depends on your use case. What do you want to achieve?

Please see the link in the original comment for code sample.

I meant: What is your use case? In words, not in code.
And sure, improving the docs would be great. Any volunteers?

Note that HistGradientBoostingRegressor and HistGradientBoostingClassifier might be better suited than the old GBDTs for large sample sizes. There, the (inverse) link functions are documented.

the code allows to see the loss as a function of stage, which lets me figure out the optimal number of estimators.

Also, I see nothing like _loss in HistGradientBoostingClassifier - except maybe that since the only loss that is supported is log_loss, my code would work there too.

the code allows to see the loss as a function of stage, which lets me figure out the optimal number of estimators.

Note that this would only give the training error. A better strategy is looking at the test/validation error, preferably with cross validation, e.g. with GridSearchCV.

the code allows to see the loss as a function of stage, which lets me figure out the optimal number of estimators.

Note that this would only give the training error. A better strategy is looking at the test/validation error, preferably with cross validation, e.g. with GridSearchCV.

I am passing test_X to staged_decision_function, not train_X.

preferably with cross validation, e.g. with GridSearchCV.

GridSearchCV is way too expensive compared to my method.